acceptodds
Under review as a conference paper at ICLR 2027

Learning Commute-Time-Preserving World Models for Planning

Abstract

World models allow agents to plan in latent space by choosing a sequence of actions that most reduces the distance to a given goal state. Thus, planning can benefit from latent representations whose distances mirror commute-times in the environment. The spectral embedding space of the graph Laplacian provides such a representation, if it obeys a specific eigenvalue-dependent scaling. Unfortunately, instantiating the graph Laplacian is intractable in large, continuous environments. Self-supervised learning offers a natural route to such commute-time-preserving embeddings at scale. However, here we show that existing methods, which commonly encourage isotropic representations to prevent representational collapse, tend to degrade the “correct” eigenvalue-dependent scaling, leading to an inaccurate representation of commute times. To address this problem, we introduce a Commute-Time-Preserving World Model (CTWM), combining a latent displacement predictor and a log-determinant regularizer that prevents collapse, which provably recovers the correctly scaled Laplacian representation under deterministic dynamics and at the predictor’s fixed point. In numerical simulations, CTWM matches or outperforms LeWM, a strong task-agnostic baseline, on several complex, continuous goal-reaching benchmarks, while using only half the parameters.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
planning, world models, self supervised learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_LHa9OpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Commute-Time-Preserving World Models (CTWM), a JEPA-based reconstruction free world model that jointly learns an encoder and latent dynamics model. The main motivation of the paper is that the latent spaces of prior JEPA-based world modeling works do not accurately reflect the commute times in the actual environment. While the scaled spectral embedding of the transition graph’s Laplacian has this property but is intractable, and thus this motivates the use of self-supervised learning methods to learn these embeddings. They argue that prior works usually incorporate regularization terms that encourage isotropic representations which inherently removes eigenvalue-dependent scaling, which is a necessary component for representing commute times properly.

The proposed architecture builds upon LeWM with residual prediction and a log-determinant regularization term. The residual predictor is trained to predict the difference between successive embeddings and incorporates a regularizer term to penalize large predictor outputs. The goal of the residual predictor to predict the displacement between latents informative and predictive while keeping temporally correlated latents close together. The log-determinant regularization term helps prevent global collapse. Prior works have opted to use methods like VCReg yet these methods impose a hard threshold on feature variance which limits their ability to represent commute-time accurately.

The main contributions of the paper are proposing their CTWM architecture that is based on LeWM/JEPA world modeling as well as providing theoretical foundations on why different regularizers may result in different representations, and more specifically, how log-determinant regularization is consistent with scaled Laplacian representation unlike VCReg and orthonormality constraints.

To validate their direction, they do a small scale empirical experiment to verify the regularization choice as well as test on 6 different goal-reaching continuous RL environments. On the small scale setting, they create a torus with MNIST-rendered states, where CTWM matches the ground truth and ALLO subspaces at the proposed ceiling and beats VICREG, and their orthonormal + residual prediction ablation. On the goal-reaching benchmarks, CTWM outperforms or at least matches the performance of LeWM on all experiments. An analysis on a discretized version of the pointmaze benchmark further suggests that CTWM's latent distances follow commute-time geometry better than LeWM.

Critical strengths

The authors present strong motivation for learning reconstruction free world models such that the latent geometry is useful for planning. This is a highly relevant area of research as prior reconstruction free/JEPA-based works have often been advocated for their efficient planning in terms of time and performance. The authors' research direction is well explained and supported with clean theoretical formulations that explain the role of the regularizer as well as provide evidence on why a log-determinant regularizer is more optimal for planning than the choices used in prior works. The authors also present 3 separate experiments sections that each answer different questions such as importance of the regularizer choice, overall success rate performance relative to LeWM on goal reaching tasks, and whether CTWM faithfully represents commute-time geometry.

Critical weaknesses

W1. The paper's main claim is that CTWM outperforms the prior work because it preserves commute times while prior regularizers for JEPA models do not. However, the experiments do not cleanly establish this link because the regularizer test uses a different baseline than the planning test. The torus experiment that checks whether each regularizer recovers the correct laplacian subspace compares CTWM against ALLO, VICreg and the their WM-GDO ablation. However, in the planning experiments CTWM is only compared against LeWM which uses SIGreg and not VICreg. Since they are both different regularization objectives, the torus results do not show that SIGreg fails to recover the correct scaling and it also remains unclear if a VICreg-based JEPA world model such as PLDM would perform on the goal reaching experiments. This matters because in the original LeWM paper, PLDM outperforms LeWM on Two-room, which is one of the environments where CTWM's long-horizon gains are larger. Additionally, there is no ablation of the slowness term introduced through residual prediction. While WM-GDO shows that replacing log-determinant with a different constraint degrades the latent geometry, no experiment tests log-determinant without the slowness penalty or what happens when the penalty is applied directly to the encoder. The paper does not clearly show that residual prediction is necessary or how much the slowness term contributes in practice.

W2. There are no comparisons to other planning methods which limits the paper's claims. LeWM is the only baseline for the goal reaching experiments, but as highlighted above, LeWM sometimes underperforms other similar works. The paper also does not compare against related methods that target improving the planning of JEPA-based models. For example, Temporal Straightening was cited and seems to be close to CTWM but is not compared against. Comparisons against temporal straightening would be a valuable baseline and other works would significantly reinforce the paper's claims that commute-time preserving latent representations are necessary.

Questions for authors

Q1. Why was the latent dimension reduced from 192 to 64? I am curious on if this choice helps CTWM's performance or CTWM outperforms LeWM despite the reduction in size. There wasn't any argument for this reduction but it is a big reason on why CTWM can claim half the parameter count. Are there any results on CTWM at D = 192 and LeWM at D = 64 to justify the statement CTWM outperforms LeWM at half the parameter count?

Q2. Why does CTWM have pretty much flat curves based on epochs across all the six goal reaching tasks? I find it a bit weird that CTWM pretty much performs the same regardless of how many epochs. The performance delta from epoch 1 versus epoch 10 looks below 10% on all tasks and virtually 0 on Cube, Two-Room, and Pointmaze. There isn't a mention on whether CTWM results in faster learning over LeWM or something else is going on. Could you analyze or give an explanation on what is happening at epoch 1 and the early epochs for there to be much flatter curves than LeWM?

Q3. What happens with log-determinant if there slowness penalty is higher or nonexistent? Alternatively, what happens when you apply a slowness penalty to LeWM?

Q4. Why do you route the slowness penalty through the predictor rather than applying it directly to the encoder? Is there an empirical difference?

Reviewer_MYigOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Commute-Time-Preserving World Models (CTWM), a reconstruction-free joint-embedding world model designed to preserve commute-time distances for goal-conditioned planning. Building on LeWM, CTWM trains the predictor to output action-conditioned latent displacements with a norm penalty and replaces SIGReg with a log-determinant covariance regularizer. The paper shows that, at the predictor's fixed point under deterministic dynamics, the encoder recovers the eigenvalue-scaled Laplacian representation.

Critical strengths
  • The paper gives a clear theoretical analysis showing how log-determinant regularization recovers the eigenvalue scaling needed for commute-time distances, unlike orthonormality or VCReg.
  • The evaluation covers a synthetic torus graph and six pixel-based continuous control tasks. CTWM matches or exceeds LeWM with half the parameters, with large gains on Cube, Scene, and Pointmaze.
Critical weaknesses
  • The conclusion claims that commute time "depends only on the environment, not on any policy" (lines 483–484). However, Theorem 1 defines the Laplacian over the dataset's transition and stationary distributions (line 918), so the learned metric depends on the behavior policy that collected the data.
  • The paper assumes reversible dynamics in its contributions and conclusion, but does not discuss how symmetric commute times apply to the irreversible manipulation tasks it evaluates (PushT, Cube, Scene).
  • There are several typos and formatting errors: "destinction" (line 248), "the underlying reasons is" (line 249), "the the" (line 316), "his is" (line 367), a broken reference to "Appendix 3" (line 506), and unremoved template text in the Acknowledgments (line 515).
Questions for authors
  • How does the learned commute-time geometry change if the same environment is trained on data from different behavior policies (e.g., random vs. expert), and what are the limitations of using a symmetric metric on irreversible manipulation tasks?
  • Why does CTWM's success rate on Reacher increase from 82.3% on 25-step goals to 98.2% on 100-step goals in Table 3?
Reviewer_U6AcOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Recent work on JEPA architectures (Le World Model for example), make use of light regularization to prevent collapse, enforcing the representation space to be isotropic gaussian. However, while useful to prevent collapse, such a representation may not be ideal for planning. The laplacian representation is potentially a better idea for the representation space, since embeddings obey distance relationships, whereby embedding distances are proportional to how far apart states are (specifically, a quadratic relationship between embedding distance and commute time). This paper explores if we can regularize a JEPA in such a way as to induce a properly scaled laplacian representation, thus getting both robust JEPA training, and superior downstream performance for planning.

Critical strengths

The central question is well motivated. Many recent works have begun to look beyond isotropic gaussian for better representation distributions, and laplacian representation makes sense from the perspective of planning.

The results seem promising. Figure 3 shows CTWM consistently beating LWM on 6 benchmark tasks, with a 25 step planning horizon.

We would expect the gap to widen as we increase goal distance, and the authors report this as well (line 428). Closer inspection shows that this claim is slightly overstated in the paper, as in appendix E.1, the gap actually narrows on one task (Cube), but overall the trend seems positive.

Critical weaknesses

In the Torus toy example, LeWM is missing, despite it being the main baseline in the rest of the paper. The authors motivate this by saying they include vicreg as “representative example of non-residual world-modeling approaches”, but this is a strange choice and sheds doubts on the conclusions of figure 2.

The paper is also missing proper ablations (outside the toy example). The author’s method introduces several changes:

  1. regularizer,
  2. residual predictor, a smaller predictor
  3. LayerNorm and no weight decay. It would be best to ablate these to isolate that indeed the central result is from the regularizer and not another change.

Weak baselines: In figure 3, in many examples, the leworld model performance is still going up, and training seems to have stopped early. While it is still noteworthy that their method converges faster, that is a different story altogether, and the paper should demonstrate that the performance gap is not simply due to undertraining.

Questions for authors
  1. Can you elaborate on: were the leworld model baselines properly saturated?
  2. Can you add Sigreg as a baseline to the torus toy example?
  3. Can you add an ablation as described above?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.